报告题目一：Efficient Resource Scheduling for Machine Learning Clusters
Chuan Wu received her B.Engr. and M.Engr. degrees in 2000 and 2002 from the Department of Computer Science and Technology, Tsinghua University, China, and her Ph.D. degree in 2008 from the Department of Electrical and Computer Engineering, University of Toronto, Canada. Between 2002 and 2004, She worked in the Information Technology industry in Singapore. Since September 2008, Chuan Wu has been with the Department of Computer Science at the University of Hong Kong, where she is currently an Associate Professor and serves as an Associate Head of the department, responsible for curriculum and development matters. Her current research is in the areas of cloud computing, distributed machine learning/big data analytics systems, network function virtualization, and data center networking. She is a senior member of IEEE, a member of ACM, and served as the Chair of the Interest Group on Multimedia services and applications over Emerging Networks (MEN) of the IEEE Multimedia Communication Technical Committee (MMTC) from 2012 to 2014. She is an associate editor of IEEE Transactions on Cloud Computing, IEEE Transactions on Multimedia, and ACM Transactions on Modeling and Performance Evaluation of Computing Systems. She has also served as TPC members and reviewers for various international conferences and journals. She was the co-recipient of the best paper awards of HotPOST 2012 and ACM e-Energy 2016.
Machine learning workloads are common in today’s production clusters due to the proliferation of AI-driven services. Efficient resource scheduling is the key to the maximal performance of a machine learning cluster. Existing cluster schedulers typically specify a fixed amount of resources for each job, prohibiting high resource efficiency and job performance. In this talk, I will share our recent work on designing a customized job scheduler for machine learning clusters (running various training jobs), which minimizes training completion time by maximising exploiting available resources. Our scheduler uses online fitting to predict model convergence during training, and sets up performance models to accurately estimate training speed as a function of allocated resources in each job. Based on the models, a simple yet effective method is designed and used for dynamically allocating resources and placing deep learning tasks to minimize job completion time. We implement our scheduler on Kubernetes, and experiments in a deep learning testbed show that our schedulers outperform existing representative cluster schedulers significantly in terms of training job completion time and resource utilisation efficiency.
报告题目二：Learning Towards Better Accuracy and Privacy
Liyao Xiang is an Assistant Professor at John Hopcroft Center for Computer Science of Shanghai Jiao Tong University since September 2018. She obtained her Ph.D. and master's degrees in the Department of Electrical and Computer Engineering at University of Toronto, respectively in 2018 and 2015. Her research interests include security and privacy in machine learning, privacy analysis in data mining, and mobile computing. Currently, she is interested in designing mechanisms to guard machine learning models against malicious attacks.
This talk starts with a real-world issue: to provide indoor localization services to satisfy contextual and ephemeral needs, e.g., at conferences or exhibitions events. We design, implement, and evaluate Tack, a new mobile application framework that uses a combination of known landmark locations, contacts over Bluetooth Low Energy, crowdsourcing, and dead-reckoning to estimate and refine user locations. At its core, an inference algorithm is designed to run on mobile devices to make the estimation more accurate.
Accuracy and privacy pose as a pair of contradictory requirements in machine learning frameworks -- stricter privacy guarantee is always achieved with degraded learning accuracy -- and such degradation is even worse with deep learning. We found the fundamental cause is that a loose characterization of utility and privacy leads to over-distortion of the model. By recognizing the accuracy-privacy tradeoff as a utility maximization problem subject to a set of privacy constraints, we lower-bound the distortion, and significantly improve the learning accuracy as compared to the state-of-the-art, under the same privacy guarantee.